Skip to content

Status: Needs Review

This page has not been reviewed for accuracy and completeness. Content may be outdated or contain errors.

How-To: Build Pipelines In YAML

Use YAML when you want a reproducible, shareable pipeline definition that can be composed into a trainrun or resolved remotely over gRPC.

Minimal Current Pipeline

metadata:
  name: MinimalPipeline
  description: Current YAML example
  author: cuvis.ai

nodes:
  - name: data_node
    class_name: cuvis_ai.node.data.LentilsAnomalyDataNode
    hparams:
      normal_class_ids: [0, 1]

  - name: rx
    class_name: cuvis_ai.anomaly.rx_detector.RXGlobal
    hparams:
      num_channels: 61
      eps: 1.0e-06

connections:
  - source: data_node.outputs.cube
    target: rx.inputs.data

Schema Rules

  • nodes is a list.
  • Each node uses name, class_name, and optional hparams.
  • Each edge uses source and target.
  • Port names use node.outputs.port and node.inputs.port.

Example: Tracking-Oriented Pipeline

metadata:
  name: TrackingOverlay
  description: Render overlays from tracking JSON
  author: cuvis.ai

nodes:
  - name: tracks
    class_name: cuvis_ai.node.json_file.TrackingResultsReader
    hparams:
      json_path: tracking_results.json

  - name: overlay
    class_name: cuvis_ai.node.anomaly_visualization.TrackingOverlayNode

  - name: writer
    class_name: cuvis_ai.node.video.ToVideoNode
    hparams:
      output_video_path: tracking_overlay.mp4

connections:
  - source: tracks.outputs.image
    target: overlay.inputs.image
  - source: tracks.outputs.bboxes
    target: overlay.inputs.bboxes
  - source: overlay.outputs.image
    target: writer.inputs.image

Compose It Into A Trainrun

# @package _global_
defaults:
  - /pipeline/sam3@pipeline: sam3_text_propagation
  - /data@data: tracking_cap_and_car
  - /training@training: default
  - _self_

name: sam3_text_demo
output_dir: ./outputs/${name}

Validation Checklist

  • Confirm the class path matches a real importable node class.
  • Confirm every required input port has a producer.
  • Prefer current checked-in configs as templates instead of historical examples.
  • Save narrow plugin manifests next to the trainrun when a pipeline depends on plugins.